Machine learning-enabled inverse design of bimaterial thermoelastic lattice metamaterials

1. Problem Statement

Thermoelastic metamaterials, particularly those based on bimaterial hybrid-honeycomb structures, offer unique capabilities such as negative thermal expansion (NTE) and negative Poisson's ratios (auxeticity). These properties are highly desirable for precision instruments and temperature-adaptive engineering systems. However, designing these materials presents two major challenges:

• Forward Prediction Complexity: Accurately predicting effective thermoelastic properties (Young's modulus, Poisson's ratio, thermal expansion) for a given complex microstructure requires computationally expensive finite element (FE) simulations and involves establishing nonlinear correlations between design variables and properties.
• Inverse Design Difficulty: Identifying the specific geometric and material parameters required to achieve a target set of properties is a non-trivial optimization problem. It is often ill-posed (non-unique solutions exist) and demands sophisticated algorithms that are slow and difficult to implement for practical engineering applications.

The authors aim to overcome these bottlenecks by developing a Machine Learning (ML)-enabled framework that facilitates both rapid forward prediction and efficient inverse design of bimaterial hybrid-honeycomb lattices.

2. Methodology

The proposed framework integrates computational homogenization with Artificial Neural Networks (ANNs) in a two-stage process:

A. Data Generation

• Structure: A 2D bimaterial hybrid-honeycomb unit cell is defined by six design variables:
  • Geometric: Aspect ratio ($h/w$), strut angles ($\theta_1, \theta_2$), and relative density ($\rho$).
  • Material: Ratios of Young's moduli ($E_I/E_N$) and thermal expansion coefficients ($\alpha_I/\alpha_N$) between inclined and non-inclined struts.
• Simulation: A large dataset of $10^5$ structures was generated by randomly sampling these variables within feasible ranges.
• Homogenization: Effective thermoelastic properties were calculated using computational homogenization via FE simulations (Timoshenko beam elements with periodic boundary conditions). This yielded a dataset pairing structural features ($S$) with effective properties ($P$).

B. Forward Prediction Model

• Architecture: A standard ANN was trained to map structural features ($S$) to effective properties ($P$).
• Preprocessing: Inputs and outputs were normalized (min-max scaling), with logarithmic transformations applied to modulus and expansion ratio inputs to handle wide dynamic ranges.
• Objective: Minimize the Mean Square Error (MSE) between predicted and ground-truth properties.

C. Inverse Design Models

• Approach: The authors employed a direct inverse design method using ANNs to map target properties ($P$) back to structural features ($S$).
• Loss Function Strategy: To address the non-uniqueness problem (where multiple designs yield the same properties), a composite loss function was used:
  $$L_{id} = \|F_{fp}(F_{id}(P_i)) - P_i\|^2 + \lambda \|F_{id}(P_i) - S_i\|^2$$
  • The first term is the reconstruction loss: The predicted structure is fed back into the trained forward model ($F_{fp}$) to check if it reproduces the target properties.
  • The second term is the prediction loss: It penalizes deviation from the training data distribution (active only during early training with $\lambda=0.2$ to aid convergence, then set to 0).
• Scenario Adaptation: Six distinct inverse models were developed to handle different practical constraints:
  1. Full Input/Full Output: All 6 properties $\to$ All 6 design variables.
  2. Partial Input/Full Output: Targeting specific properties (e.g., NTE + Auxeticity) $\to$ All 6 variables.
  3. Fixed Material/Full Output: Fixed base materials $\to$ Only geometric variables.
  4. Fixed Material/Partial Output: Fixed materials + specific targets $\to$ Geometric variables.

3. Key Contributions

1. Hybrid ML Framework: The development of a robust pipeline combining high-throughput FE simulation with deep learning for both forward and inverse tasks.
2. Reconstruction Loss Mechanism: The specific implementation of the forward model within the inverse model's loss function effectively resolves the non-uniqueness issue without requiring generative models, ensuring the designed structure actually possesses the target properties.
3. Versatile Inverse Models: The creation of six specialized models covering diverse engineering scenarios, including cases where base materials are pre-selected (fixed) and cases where only a subset of target properties is specified.
4. Conditional Optimization Capability: The framework enables the determination of attainable property limits (min/max) for a fixed parameter, acting as a powerful tool for constrained optimization.

4. Results

• Forward Prediction Performance: The forward ML model achieved high accuracy with coefficients of determination ($R^2$) greater than 0.994 for all six effective thermoelastic constants, demonstrating near-perfect agreement with FE simulations.
• Inverse Design Accuracy:
  • In Case 1 (Full design), the models successfully predicted structures matching target properties with relative errors ranging from 3.4% to 6.4%.
  • In Cases 2–4 (Partial targets or fixed materials), the models maintained high accuracy (relative errors generally < 5%).
  • The models successfully identified distinct structural configurations that yielded the same target properties, confirming their ability to navigate the non-unique solution space.
• Conditional Search: The iterative search method successfully determined the minimum and maximum achievable thermal expansion coefficients for a fixed Poisson's ratio (and vice versa), validating the models' utility for exploring design spaces.

5. Significance and Conclusion

This work demonstrates that machine learning can significantly accelerate the design cycle of complex thermoelastic metamaterials. By replacing iterative, computationally heavy optimization algorithms with trained neural networks, the authors provide a tool that:

• Reduces Computational Cost: Inverse design becomes instantaneous compared to traditional optimization loops.
• Enhances Practicality: The ability to fix base materials (a common real-world constraint) and still find optimal geometries makes the method directly applicable to manufacturing.
• Facilitates Discovery: The models can rapidly identify feasible design regions and property limits, aiding engineers in creating custom thermal expansion and mechanical response components for aerospace and precision engineering.

The proposed framework is generalizable and can be extended to other types of thermoelastic metamaterials, marking a significant step toward the data-driven design of advanced functional materials.